When Two Tech Giants Agree on Something, You Should Probably Listen

“We’ve arrived at the era of useful AI”

That’s a direct Huang quote, and the line that stuck with me most. For years, AI felt like this perpetually impressive but slightly impractical thing. Great at generating images, decent at writing emails, occasionally wrong in ways you couldn’t predict.

But Huang’s framing at Dell Technologies World was different. He described how agentic AI — AI that doesn’t just respond but actually acts — has fundamentally changed what computing even means. Old model: human asks question, AI answers. New model: human sets a goal, AI figures out the steps, uses tools, accesses memory, and gets it done.

If you’ve ever set up an AI agent pipeline — LangChain, AutoGPT, newer enterprise tooling — this maps directly to what you’ve experienced. The model alone isn’t the agent. The scaffolding around it is what makes it actually work.

The memory problem nobody was talking about loudly enough

Here’s the part I found most interesting — and honestly a bit alarming for anyone planning infrastructure budgets.

The reason is agentic AI. When an agent is running — pinging databases, checking its own memory for context, executing tools, running inference — it’s hitting memory constantly and fast. Huang noted that agent frameworks require far more CPUs than anything we built for before, because agents use tools constantly compared to how slowly human users interact.

Dell put a number on the shift: companies that fully reimagine their workflows with AI agents aren’t seeing 10–30% efficiency gains. They’re seeing 10x to 30x improvements. Ten to thirty times. That kind of jump changes what hardware you need, how much of it, and how fast.

I’ve run into this personally at a smaller scale. Testing local AI agents on my MacBook earlier this year, I kept hitting walls — not because my GPU was maxed, but because I’d run out of fast memory for the context window. Scale that to an enterprise running thousands of agents simultaneously, and you understand why memory is now the central hardware conversation.

“Demand is going parabolic. Utterly parabolic.”

Huang wasn’t being hyperbolic for effect. He said computation demand for agentic AI has grown 100x to 1,000x depending on the workload — not since five years ago, but in the recent past.

The supply chain is already strained. Dell added 1,000 new AI server clients in a single quarter. Huang acknowledged the semiconductor supply chain is growing — but demand is growing faster. Michael Dell described a chain so intricate it almost sounds like an engineering marvel in itself: silicon photonics, advanced CPUs, GPUs, high-bandwidth memory, all needing to be sourced and assembled at a pace the industry has never sustained before.

The jump from Blackwell to Rubin is the kind of cost inflection that changes what’s economically viable. Things that cost too much to run six months ago might be deployable now.

The China question: more nuanced than the headlines

Huang spoke days after joining President Trump’s summit in China. His statement — that he expects China’s market will eventually open up to US AI chips — grabbed headlines. But the actual quote is more careful:

That’s not triumphalism. That’s reading the situation honestly. Huang confirmed Nvidia is committed to complying with existing regulations — he’s not predicting policy change because he wants it, he’s observing that China’s AI sector needs high-performance chips, and protectionism has its own costs.

Michael Dell, who recently returned from China himself, added that the US administration wants to ensure American leadership in AI — meaning export restrictions exist for a reason, but so does market access. It’s a tension that doesn’t resolve cleanly. Dell’s emphasis on Taiwan as a vital manufacturing partner tells you this is a multi-year geopolitical chess game, not a single move.

Personal AI: the part that got underplayed

Huang’s concept of “personal AI” deserves more attention — the shift from centralized cloud AI to AI running at the point of context: on your laptop, inside a factory machine, inside a vehicle.

Michael Dell backed this up practically: Dell is embedding the ability to run small local models directly inside PCs. Not as a gimmick — but because for knowledge work, the PC is still where people actually do things. If your agent needs to access sensitive company data, you don’t want it routing through a public cloud. Confidential computing on-device is the answer.

I’ve been testing this with a smaller team at work — using a local LLM for tasks involving proprietary data. The performance gap between cloud and local has closed more than most people realize. Not for everything, but for a surprising range of tasks.

What mistakes are people making right now?

Both CEOs were speaking broadly about enterprise adoption. Here’s what I’ve seen from the trenches — because there are real mistakes companies make as they rush into agentic AI.